Content preference crowd selection method based on causal inference

Abstract

The invention belongs to the technical field of causal inference, and discloses a causal inference-based content preference crowd selection method, which comprises the following steps of: A1, selecting a user group; a2, dividing user groups; a3, a Persuadable crowd is located, and the Persuadable crowd is located; and A4, obtaining a UPLift value. The invention provides an innovative crowd selection scheme which is specially designed for coping with a complex and changeable new scene, the scheme can reasonably optimize resources from the global perspective, and meanwhile, the scheme also has certain self-optimization capability, continuously improves the selection effect, and is suitable for popularization and application. The scheme also plays an important role in making strategies such as content recommendation, search and push in a support platform, the platform can more effectively push contents interested by users and improve user experience and satisfaction through accurately delineating user groups, in addition, the scheme also provides powerful support for accurate marketing and refined operation, and the user experience and satisfaction are improved. And the platform is helped to realize more efficient user conversion and income increase.

Description

Technical Field

[0001] The present invention belongs to the technical field of causal inference, and specifically is a method for selecting a content preference group based on causal inference. Background Art

[0002] The current automotive industry is mainly divided into two categories based on the power type: one is traditional fuel vehicles, and the other is new energy vehicles covering pure electric, hybrid and other modes. The content ecology of the Autohome platform is built around these two categories. For fuel vehicles, Autohome already has a stable user base, and its content has a high click-through rate and excellent performance in the field of fuel vehicles; as for new energy vehicles, it represents the future development trend of the automotive industry, so Autohome needs to increase resource investment and make it a key development direction.

[0003] In actual business operations, when recommending content on the homepage, we first need to segment users and identify user groups with a high interest in new energy vehicles in order to optimize the display of new energy content and increase their click-through rate and browsing time. However, if we directly rely on response models (such as new energy preference models, click-through rate (Ctr) models, retention intention models, etc.) for screening, we may screen out user flows that already have high click-through rates. These users tend to click and leave information no matter what content is pushed. This situation will bring two problems: First, it may squeeze the display space for fuel vehicle content, resulting in a decrease in clicks on fuel vehicle content, damaging the established fuel vehicle content performance indicators; second, the overall content recommendation effect improvement may be limited, because the user group with a high basic click-through rate limits the space for further growth, which greatly reduces the optimization effect. Summary of the invention

[0004] The purpose of the present invention is to provide a method for selecting content preference groups based on causal inference to solve the problems raised in the above-mentioned background technology.

[0005] In order to achieve the above-mentioned object, the present invention provides the following technical solution: a method for selecting a content preference group based on causal inference, the steps of the method are as follows:

[0006] A1, select user groups: Based on the proposed UPLift model, select user groups that will only click when they see content related to new energy vehicles;

[0007] A2, dividing user groups: users are divided into four categories based on intervention and click conditions;

[0008] A3, locate the Persuadables population: use the UPLift gain model to perform counterfactual predictions to calculate causal effects and identify the Persuadables population;

[0009] A4: Obtain the UPLift value based on the T-learner model and the S-learner model, and perform CTR calculation.

[0010] Preferably, the four categories in A2 are:

[0011] Sure Things: Clicks will come without intervention;

[0012] Persuadables: No intervention, no click, only click after intervention;

[0013] Do Not Disturbs: If you do not interfere, you will get points, but if you interfere, you will get no points;

[0014] Lost Causes: Not selected regardless of intervention.

[0015] Preferably, the method based on the T-learner model in A4 is:

[0016] B1, machine learning: in the new energy crowd scenario, use machine learning as the base learner to estimate the gain;

[0017] B2, Classification: Label according to content and divide into new energy vehicles and fuel vehicles;

[0018] B3, construct ML model: construct ML models for the experimental group and the control group respectively, and perform calculations;

[0019] B4, obtain intervention gain value: after obtaining two groups of ML models, exchange the prediction samples, and after obtaining the scores, subtract the scores to obtain the intervention gain value;

[0020] Preferably, the S-learner-based model in A4 first constructs treatment as a feature. In the new energy population scenario, after the model training is completed, counterfactual reasoning is performed by changing the treatment variable to calculate the uplift value of the sample after the intervention.

[0021] Preferably, during the model training process, on the one hand, it is necessary to avoid overfitting of the model, and on the other hand, it is necessary to ensure that the model is fully trained.

[0022] Preferably, the estimated gain is:

[0023] 1.μ1(x)=E[Y / T=1,X=x];μ0(x)=E[Y / T=0,X=x];

[0024]

[0025] In formula 1, the response model μ1(x) is used to fit the relationship between the response target Y and the feature X when the intervention is applied, while the response μ0(x) is used to fit the relationship between the response target Y and the feature X when the intervention is not applied;

[0026] Formula 2 is modified to T = μ1(x)-μ0(x), which represents the gain brought to the user after intervention on a single sample x in the test set.

[0027] Preferably, the S-learner model is characterized by:

[0028] 1.μ(t,x)=E[Y|T=t,X=x];

[0029] 2.T=u(1,x)-u(0,x),.

[0030] Preferably, during the model training process, the parameter values ​​of the two models are as similar as possible to reduce the accumulation of deviations at the model level.

[0031] The beneficial effects of the present invention are as follows:

[0032] The present invention proposes an innovative crowd selection scheme, which is specially designed to cope with complex and changeable new scenarios. The scheme can reasonably optimize resources from a global perspective to ensure that each resource can play the maximum effect. The method has clear and intuitive logic, which is not only easy to implement, but also has excellent explainability, so that relevant personnel can easily understand and apply it. By introducing user preference data, the scheme can further improve the accuracy of selection and ensure that the selected user group is highly matched with the marketing goal. At the same time, the scheme also has a certain self-optimization ability, and can be dynamically adjusted according to the actual operation effect to continuously improve the selection effect. The scheme also plays an important role in supporting the formulation of content recommendation, search, push and other strategies within the platform. By accurately defining the user group, the platform can more effectively push the content that users are interested in, improve user experience and satisfaction. In addition, the scheme also provides strong support for precision marketing and refined operations, helping the platform to achieve more efficient user conversion and revenue growth. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 It is a schematic diagram of the process of the present invention;

[0034] Figure 2 Schematic diagram of the T-learner model method of the present invention. DETAILED DESCRIPTION

[0035] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0036] like Figure 1 to Figure 2 As shown, an embodiment of the present invention provides a method for selecting a content preference group based on causal inference, and the steps of the method are:

[0037] A1, select user groups: Based on the proposed UPLift model, select user groups that will only click when they see content related to new energy vehicles;

[0038] A2, dividing user groups: users are divided into four categories based on intervention and click conditions;

[0039] A3, locate the Persuadables population: use the UPLift gain model to perform counterfactual predictions to calculate causal effects and identify the Persuadables population;

[0040] A4: Obtain the UPLift value based on the T-learner model and the S-learner model, and perform CTR calculation.

[0041] By proposing an innovative crowd selection solution, which is specially designed to cope with complex and changing new scenarios, the solution can optimize resources reasonably from a global perspective to ensure that every resource can play the maximum effect. The method is clear and intuitive in logic. It is not only easy to implement, but also has excellent explainability, so that relevant personnel can easily understand and apply it. By introducing user preference data, the solution can further improve the accuracy of selection and ensure that the selected user group is highly matched with the marketing goal. At the same time, the solution also has a certain self-optimization ability, and can be dynamically adjusted according to the actual operation results to continuously improve the selection effect. The solution also plays an important role in supporting the formulation of content recommendation, search, push and other strategies within the platform. By accurately defining the user group, the platform can more effectively push content that users are interested in, improve user experience and satisfaction. In addition, the solution also provides strong support for precision marketing and refined operations, helping the platform to achieve more efficient user conversion and revenue growth.

[0042] Among them, the four categories in A2 are:

[0043] Sure Things: Clicks will come without intervention;

[0044] Persuadables: No intervention, no click, only click after intervention;

[0045] Do Not Disturbs: If you do not interfere, you will get points, but if you interfere, you will get no points;

[0046] Lost Causes: Not selected regardless of intervention.

[0047] By dividing the categories, we can identify the key points. For Sure Things users, they have a high willingness to click, and the solution can ensure the effective use of resources and achieve high conversion rates without too much intervention. For Persuadables users, the solution can effectively improve their click-through rate through precise intervention, thereby tapping potential value. For Do Not Disturbs users, the solution can intelligently identify and avoid unnecessary disturbances, ensuring that they maintain a high level of activity in a natural state and avoiding resentment caused by excessive intervention. For Lost Causes users, the solution can reduce the investment in invalid resources, focus more on valuable users, and improve overall operational efficiency.

[0048] Among them, the method based on the T-learner model in A4 is:

[0049] B1, machine learning: in the new energy crowd scenario, use machine learning as the base learner to estimate the gain;

[0050] B2, Classification: Label according to content and divide into new energy vehicles and fuel vehicles;

[0051] B3, construct ML model: construct ML models for the experimental group and the control group respectively, and perform calculations;

[0052] B4, obtain intervention gain value: after obtaining two groups of ML models, exchange the prediction samples, and after obtaining the scores, subtract the scores to obtain the intervention gain value;

[0053] By using machine learning as the base learner and training the model with rich historical data, we can accurately estimate the gain of new energy content on user behavior. This step not only improves the accuracy of the prediction, but also lays a solid foundation for subsequent steps. Next, we label the content according to its characteristics, clearly distinguish between new energy and fuel vehicles, and ensure that subsequent model construction and calculation can be targeted. Subsequently, we build machine learning models for the experimental group and the control group respectively. By comparing the performance of the two groups of models on the same data set, we can more accurately capture the impact of intervention measures on user behavior. Finally, by exchanging prediction samples and calculating scores, we can obtain the prediction results of the two groups of models under different conditions. By subtracting the two groups of scores, we can get the intervention gain value, thereby quantitatively evaluating the effect of intervention measures on users' willingness to click on new energy content.

[0054] Among them, the S-learner-based model in A4 first constructs treatment as a feature. In the new energy population scenario, after the model training is completed, counterfactual reasoning is performed by changing the treatment variable to calculate the uplift value of the sample after intervention.

[0055] By incorporating treatment as a feature into model training, the model can learn the impact of intervention on user behavior. After model training is completed, by changing the value of the treatment variable, we can perform counterfactual reasoning, that is, simulate the user's behavior under different intervention conditions. This step is crucial for evaluating the effectiveness of intervention measures. Specifically, we can calculate the uplift value of the sample after the intervention, that is, the degree to which the intervention improves the user's willingness to click on new energy content. This quantitative indicator not only helps us accurately evaluate the effect of the intervention, but also provides strong support for subsequent precision marketing and refined operations.

[0056] During the model training process, on the one hand, it is necessary to avoid overfitting of the model, and on the other hand, it must be fully trained.

[0057] In the process of model training, balancing overfitting and sufficient training is a crucial step. Overfitting will cause the model to perform well on the training data, but poor generalization ability on new data, while insufficient training will make the model unable to capture the real laws of the data, affecting the prediction effect. In order to avoid overfitting, we usually use regularization techniques, data enhancement, cross-validation and other methods. These measures can limit the complexity of the model and make it more generalized. At the same time, we also need to pay close attention to the performance of the validation set during the training process. Once signs of overfitting are found, the training strategy should be adjusted in time. On the other hand, in order to ensure that the model is fully trained, we need to choose the appropriate optimization algorithm, learning rate and number of iterations so that the model can fully learn the inherent laws of the data. In addition, the training process can be further optimized through techniques such as early stopping and learning rate decay.

[0058] The estimated gain is:

[0059] 1.μ1(x)=E[Y / T=1,X=x];μ0(x)=E[Y / T=0,X=x];

[0060]

[0061] In formula 1, the response model μ1(x) is used to fit the relationship between the response target Y and the feature X when the intervention is applied, while the response μ0(x) is used to fit the relationship between the response target Y and the feature X when the intervention is not applied;

[0062] Formula 2 is modified to T = μ1(x)-μ0(x), which represents the gain brought to the user after intervention on a single sample x in the test set.

[0063] Among them, the characteristics of the S-learner model are:

[0064] 1.μ(t,x)=E[Y|T=t,X=x];

[0065] 2.T=u(1,x)-u(0,x),.

[0066] During the model training process, the parameter values ​​of the two models are kept as similar as possible to reduce the accumulation of deviations at the model level.

[0067] Ensuring that the parameter values ​​of the two models are as similar as possible is the key to reducing the accumulation of deviations at the model level. This approach helps to improve the stability and consistency of the model, so that the two models can produce similar outputs when facing the same input. By reducing the differences between the models, the prediction error caused by model deviation can be reduced, thereby improving the overall prediction performance. In addition, parameter convergence can also help simplify the model selection and tuning process, and facilitate subsequent model deployment and application. Therefore, paying attention to parameter convergence during model training is of great significance to improving model quality and prediction accuracy.

[0068] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0069] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for selecting content preference groups based on causal inference, characterized by: The steps of this method are: A1, select user groups: Based on the proposed UPLift model, select user groups that will only click when they see content related to new energy vehicles; A2, dividing user groups: users are divided into four categories based on intervention and click conditions; A3, locate the Persuadables population: use the UPLift gain model to perform counterfactual predictions to calculate causal effects and identify the Persuadables population; A4: Obtain the UPLift value based on the T-learner model and the S-learner model, and perform CTR calculation.

2. According to claim 1, a method for selecting content preference groups based on causal inference is characterized by: The A2 divides the user groups into four categories: Sure Things: Clicks will come without intervention; Persuadables: No intervention, no click, only click after intervention; Do Not Disturbs: If you do not interfere, you will get points, but if you interfere, you will get no points; Lost Causes: Not selected regardless of intervention.

3. The method for selecting content preference groups based on causal inference according to claim 1, characterized in that: The method based on the T-learner model in A4 for obtaining the UPLift value is: B1, machine learning: in the new energy crowd scenario, use machine learning as the base learner to estimate the gain; B2, Classification: Label according to content and divide into new energy vehicles and fuel vehicles; B3, construct ML model: construct ML models for the experimental group and the control group respectively, and perform calculations; B4, obtain the intervention gain value: After obtaining the two groups of ML models, exchange the prediction samples, and after obtaining the scores, subtract the scores to obtain the intervention gain value.

4. The method for selecting content preference groups based on causal inference according to claim 1 is characterized by: The A4 obtains the UPLift value based on the S-learner model. First, the treatment is constructed as a feature. In the new energy population scenario, after the model training is completed, counterfactual reasoning is performed by changing the treatment variable to calculate the uplift value of the sample after the intervention.

5. The method for selecting content preference groups based on causal inference according to claim 1 is characterized by: During model training, on the one hand, we need to avoid overfitting of the model, and on the other hand, we need to ensure that the model is fully trained.

6. A method for selecting content preference groups based on causal inference according to claim 3, characterized in that: The estimated gain is: 1.μ1(x)=E[Y / T=1,X=x];μ0(x)=E[Y / T=0,X=x]; 2. In formula 1, the response model μ1(x) is used to fit the relationship between the response target Y and the feature X when the intervention is applied, while the response μ0(x) is used to fit the relationship between the response target Y and the feature X when the intervention is not applied; Formula 2 is modified to T = μ1(x)-μ0(x), which represents the gain brought to the user after intervention on a single sample x in the test set.

7. The method for selecting content preference groups based on causal inference according to claim 1, characterized in that: The characteristics of the S-learner model are: 1.μ(t,x)=E[Y|T=t,X=x]; 2.T=u(1,x)-u(0,x),.

8. The method for selecting content preference groups based on causal inference according to claim 1, characterized in that: During the model training process, the parameter values ​​of the two models are kept as similar as possible to reduce the accumulation of deviations at the model level.

Images